Strategic Feedback Mechanisms and Experimentation

This paper studies how an information designer can influence experimentation with a new product of unknown quality. The designer commits to a history-contingent recommendation policy. A sequence of consumers arrives, not knowing their positions in the sequence or past outcomes, observing only the designer's recommendations. The designer learns from a noisy signal of quality generated from each purchase. When the designer maximizes expected consumer surplus, the first-best level of experimentation is achievable under general conditions. With alternative objectives—such as internalizing externalities on non-consumers or maximizing profit—the first best may fail, and at least one consumer incentive constraint must bind. We characterize the optimal second-best mechanisms under perfect good news and perfect bad news learning processes. With perfect good news, the designer can increase or decrease experimentation relative to the consumer-welfare benchmark until the relevant incentive constraint binds, and optimal policy may involve mixing before stopping. With perfect bad news, a designer with weaker experimentation incentives cannot do better than delaying experimentation, while a designer with stronger experimentation incentives cannot extend experimentation but may increase adoption by occasionally recommending a product already revealed to be of low quality.